jeremycochoy
Re:Eagle 7B: Soaring past Transformers
on Tuesday January 30, 2024 @05:31PM (#39196839)
Well, RWKV is using some linear (non-quadratic) form of attention so... strictly speaking... you still need a bit of attention :D
Re:Eagle 7B: Soaring past Transformers
on Tuesday January 30, 2024 @05:28PM (#39196803)
I believe RWKV is actually an architecture that can be used for encoding: given a LSTM/GRU, you can simply take the last state as an encoding of your sequence. The same should be possible with RWKV, right?
Re:Ask HN: Should We Move Again?
on Saturday November 28, 2020 @09:25AM (#25237672)
Although after doing some research I figured out what you are referencing for Hungary with some googling, some references would have been appreciated. Do you have some references about poland/hungaey, and more specifically how it can impact a business?
Re:Show HN: An algorithm automatically remove text from natural images
on Tuesday August 20, 2019 @08:52AM (#20746359)
Hello HN community. I am developing a web service which remove text from natural images, and a friend recommended me to share it with you.
A nice test is for example the doge meme from wikipedia. (https://www.google.com/url?sa=i&source=images&cd=&ved=2ahUKE...) Not every image is perfect. He is also having a hard time with huge capital fonts.
I think the usage is straight forward, but I would like to add some comments on the project.
* You are limited to 10 images per IP for now, because I have to pay for the servers and it can quickly get expensive. This is the easiest way I found to allow everybody to test it without exploding my bank account.
## Some question I am anticipating:
1) Why is it so slow
The algorithm is computationally intensive. With the current setup, it takes around 40s for a maximal size images. I could reduce it to few seconds (tested, it works) at the cost of expensive hardware.
2) Pricing page?
Yes, I sell "computation" tickets, which basically help me to compensate the cost of the servers. This is also a way to thanks me if you like this app.
3) How can I help ?
If you find bugs, please report them to me. If you find natural images on which the algorithm perform poorly, you can also share the information. If you have friends who make a good usage of this app, some visibility would definitively helps. Finally, if you want to invest or make a donation, please contact me.
If you have any question, post it below and I will answer you as best as I can.